DayFR Euro

Mistral AI drops a bomb called Pixtral-Large, capable of beating Gemini 1.5 Pro and GPT-4o

Mistral AI picks up the pace. The most prominent French start-up in the field of AI wants to prove itself. And to achieve this, it does not hesitate to go all out on the development of its models. Proof of this is with its latest model, just released: Pixtral Large. In detail, it is an open-weighted multimodal model with 124 billion parameters (just that) built on the basis of Mistral Large 2.

Second model of the family of multimodal models, it demonstrates an understanding of images of “border level“, claims the start-up, emphasizing its ability to understand documents, graphics and natural images, while retaining the cutting-edge text understanding of Mistral Large 2. It also has a context window of 128,000 tokens and can contain at least 30 high-resolution images.

Mistral competes against Anthropic, Google and OpenAI

In terms of performance, Pixtral Large breaks records. Evaluated against frontier-type models on a set of standard multimodal benchmarks, it turns out to be better than the models published by Mistral's direct competitors. Thus, on MathVista, which evaluates complex mathematical reasoning on visual data, the model achieves a score of 69.4%, outperforming all other models. By comparison, Llama-3.2 90B reaches 49.1%, Gemini-1.5 Pro 67.8%, GPT-4o 65.4% and Claude-3.5 Sonnet 67.1%.

To assess reasoning abilities on complex graphics and documents, Mistral teams relied on the ChartQA and DocVQA tests, where Pixtral Large also outperforms GPT-4o and Gemini-1.5 Pro.

Finally, Pixtral Large demonstrates competitive capabilities on the open source MM-MT-Bench test intended to reflect real-world use cases of multimodal LLMs. It outperforms Claude-3.5 Sonnet, Gemini-1.5 Pro and GPT-4o (newest). The model is available under the Mistral Research License (MRL) for research and educational use, says the start-up, adding that it is also available under the Mistral commercial license for experimentation, testing and production at commercial purposes.

Updated Mistral Large

In addition to Pixtral Large, Mistral Large, its multilingual model published last February, benefits from an update. Dedicated to high-level reasoning for complex tasks, it is now available on pixtral-large-latest, the start-up's API, and under the name Mistral Large 24.11 on Hugging Face under the Mistral Research license for research , or with a commercial license from Mistral AI for commercial use.

Compared to Large 24.07, this version benefits from improvements in understanding the long context, the addition of a system prompt and a more precise function call. “The model performs very well for RAG and agentic workflows, making it a suitable choice for enterprise use cases such as knowledge exploration and sharing, semantic document understanding, automation tasks and improving the customer experience”, comments the start-up. The model should quickly be available on supplier platforms, startingcer by Google Cloud and Microsoft Azure within a week.

The “Le Chat” interface capable of competing with ChatGPT

Mistral likes to make notable entrances. And the latest version of its conversational interface “Le Chat” is a good example of this. In its latest update, the interface benefits from numerous additions that will make OpenAI and its famous ChatGPT or even Google with Gemini pale in comparison. Latest features include: web search with citations, canvas for ideation, online editing and export, integration of the latest Pixtral Large template for better understanding of documents and images, generation of images, powered by Black Forest Labs Flux Pro.

The Canvas tool strongly resembles the interface with the eponymous name launched by OpenAI at the beginning of the month. Simply put, the interface is displayed in the chat window when the user needs to go beyond conversations and get into creation. It is possible to use Mistral's different templates on shared results and edit created content directly online without regenerating responses, creating draft versions and previewing designs.

Determined to stand out, Mistral assures that it does not seek to continue “AGI at all costs; instead, our mission is to put cutting-edge AI in your hands.” On “Le Chat”, the French flagship therefore offers a free level with these beta features and is working on the development of premium versions with higher service guarantees.

Selected for you

-

Related News :